@fugood/llama.node 0.3.15 → 0.3.17

package/src/llama.cpp/src/llama-model.cpp

@@ -17,6 +17,7 @@
 #include <cmath>
 #include <functional>
 #include <map>
+#include <regex>
 #include <sstream>
 #include <stdexcept>
 
@@ -42,11 +43,14 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_770M:          return "770M";
         case LLM_TYPE_780M:          return "780M";
         case LLM_TYPE_0_5B:          return "0.5B";
+        case LLM_TYPE_0_6B:          return "0.6B";
         case LLM_TYPE_1B:            return "1B";
         case LLM_TYPE_1_3B:          return "1.3B";
         case LLM_TYPE_1_4B:          return "1.4B";
         case LLM_TYPE_1_5B:          return "1.5B";
         case LLM_TYPE_1_6B:          return "1.6B";
+        case LLM_TYPE_1_7B:          return "1.7B";
+        case LLM_TYPE_1_8B:          return "1.8B";
         case LLM_TYPE_2B:            return "2B";
         case LLM_TYPE_2_8B:          return "2.8B";
         case LLM_TYPE_2_9B:          return "2.9B";
@@ -64,6 +68,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_15B:           return "15B";
         case LLM_TYPE_16B:           return "16B";
         case LLM_TYPE_20B:           return "20B";
+        case LLM_TYPE_27B:           return "27B";
         case LLM_TYPE_30B:           return "30B";
         case LLM_TYPE_32B:           return "32B";
         case LLM_TYPE_34B:           return "34B";
@@ -72,6 +77,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_65B:           return "65B";
         case LLM_TYPE_70B:           return "70B";
         case LLM_TYPE_236B:          return "236B";
+        case LLM_TYPE_290B:          return "290B";
         case LLM_TYPE_314B:          return "314B";
         case LLM_TYPE_671B:          return "671B";
         case LLM_TYPE_SMALL:         return "0.1B";
@@ -86,7 +92,10 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_16x3_8B:       return "16x3.8B";
         case LLM_TYPE_10B_128x3_66B: return "10B+128x3.66B";
         case LLM_TYPE_57B_A14B:      return "57B.A14B";
-        case LLM_TYPE_27B:           return "27B";
+        case LLM_TYPE_17B_16E:       return "17Bx16E (Scout)";
+        case LLM_TYPE_17B_128E:      return "17Bx128E (Maverick)";
+        case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_235B_A22B:     return "235B.A22B";
         default:                     return "?B";
     }
 }
@@ -255,7 +264,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara
     return nullptr;
 }
 
-// CPU: ACCEL -> CPU extra -> GPU host -> CPU
+// CPU: ACCEL -> GPU host -> CPU extra -> CPU
 static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices) {
     buft_list_t buft_list;
 
@@ -271,19 +280,6 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
         }
     }
 
-    // add extra buffer types
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
-    if (ggml_backend_dev_get_extra_bufts_fn) {
-        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
-        while (extra_bufts && *extra_bufts) {
-            buft_list.emplace_back(cpu_dev, *extra_bufts);
-            ++extra_bufts;
-        }
-    }
-
     // add a host buffer type
     // storing the tensors in a host buffer is useful when the processing of large batches
     // is offloaded to a GPU device, since it reduces the time spent on data transfers
@@ -298,6 +294,20 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
         }
     }
 
+    // add extra buffer types, only if no GPU device is present
+    // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
+    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+    if (ggml_backend_dev_get_extra_bufts_fn) {
+        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+        while (extra_bufts && *extra_bufts) {
+            buft_list.emplace_back(cpu_dev, *extra_bufts);
+            ++extra_bufts;
+        }
+    }
+
     // add the CPU buffer type
     for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
         ggml_backend_dev_t dev = ggml_backend_dev_get(i);
@@ -375,9 +385,12 @@ struct llama_model::impl {
     layer_dev dev_input = {};
     layer_dev dev_output = {};
     std::vector<layer_dev> dev_layer;
+
+    bool has_tensor_overrides;
 };
 
 llama_model::llama_model(const llama_model_params & params) : params(params), pimpl(std::make_unique<impl>()) {
+    pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
 }
 
 llama_model::~llama_model() {}
@@ -543,6 +556,25 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     }
                 }
             } break;
+        case LLM_ARCH_LLAMA4:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+                ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,   hparams.n_moe_layer_step);
+                hparams.n_swa_pattern = 4;    // pattern: 3 chunked - 1 full
+                hparams.n_attn_chunk  = 8192; // should this be a gguf kv? currently it's the same for Scout and Maverick
+                hparams.n_swa = 1; // TODO @ngxson : this is added to trigger the SWA branch (we store the chunked attn mask in the SWA tensor), will need to clean this up later
+
+                switch (hparams.n_expert) {
+                    case 16:  type = LLM_TYPE_17B_16E; break;
+                    case 128: type = LLM_TYPE_17B_128E; break;
+                    default:  type = LLM_TYPE_UNKNOWN;
+                }
+
+                if (type == LLM_TYPE_17B_128E) {
+                    hparams.use_kq_norm = false;
+                }
+            } break;
         case LLM_ARCH_DECI:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -667,10 +699,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
             } break;
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
                 ml.get_key(LLM_KV_ATTENTION_CAUSAL,           hparams.causal_attn);
                 ml.get_key(LLM_KV_POOLING_TYPE,               hparams.pooling_type);
+                ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS,         hparams.moe_every_n_layers, 0);
 
                 if (hparams.n_layer == 12 && hparams.n_embd == 768) {
                     type = LLM_TYPE_137M;
@@ -759,6 +793,28 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_QWEN3:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 28: type = hparams.n_embd == 1024 ? LLM_TYPE_0_6B : LLM_TYPE_1_7B; break;
+                    case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break;
+                    case 40: type = LLM_TYPE_14B; break;
+                    case 64: type = LLM_TYPE_32B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_QWEN3MOE:
+            {
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);
+
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 48: type = LLM_TYPE_30B_A3B; break;
+                    case 94: type = LLM_TYPE_235B_A22B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_PHI2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -1112,6 +1168,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
                 }
                 ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
+                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla, false);
+                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale);
@@ -1131,6 +1189,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_PLM:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
+                switch (hparams.n_layer) {
+                    case 32: type = LLM_TYPE_1_8B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_CHATGLM:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1152,6 +1219,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GLM4:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 40: type = LLM_TYPE_9B; break;
+                    case 61: type = LLM_TYPE_32B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_BITNET:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1317,6 +1393,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_GROUPNORM_GROUPS, hparams.n_norm_groups);
                 ml.get_key(LLM_KV_ATTENTION_CAUSAL,           hparams.causal_attn);
             } break;
+        case LLM_ARCH_BAILINGMOE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+
+                switch (hparams.n_layer) {
+                    case 28: type = LLM_TYPE_16B; break;
+                    case 88: type = LLM_TYPE_290B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -1544,9 +1635,26 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     GGML_ABORT("invalid layer %d for tensor %s", info.layer, tn.str().c_str());
             }
 
-            ggml_backend_buffer_type_t buft = select_weight_buft(hparams, t_meta, op, *buft_list);
+            ggml_backend_buffer_type_t buft = nullptr;
+
+            // check overrides
+            if (ml.tensor_buft_overrides) {
+                std::string tensor_name = tn.str();
+                for (const auto * overrides = ml.tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) {
+                    std::regex pattern(overrides->pattern);
+                    if (std::regex_search(tensor_name, pattern)) {
+                        LLAMA_LOG_DEBUG("tensor %s buffer type overriden to %s\n", tensor_name.c_str(), ggml_backend_buft_name(overrides->buft));
+                        buft = overrides->buft;
+                        break;
+                    }
+                }
+            }
+
             if (!buft) {
-                throw std::runtime_error(format("failed to find a compatible buffer type for tensor %s", tn.str().c_str()));
+                buft = select_weight_buft(hparams, t_meta, op, *buft_list);
+                if (!buft) {
+                    throw std::runtime_error(format("failed to find a compatible buffer type for tensor %s", tn.str().c_str()));
+                }
             }
 
             // avoid using a host buffer when using mmap
@@ -1642,6 +1750,56 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
                     }
                 } break;
+            case LLM_ARCH_LLAMA4:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Llama 4 requires n_moe_layer_step > 0");
+                    for (int i = 0; i < n_layer; ++i) {
+                        bool is_moe_layer = (i + 1) % hparams.n_moe_layer_step == 0;
+
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+                        if (is_moe_layer) {
+                            int n_ff_exp = hparams.n_ff_exp;
+
+                            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff_exp, n_embd, n_expert}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+
+                            // Shared expert
+                            const int64_t n_ff_shexp = n_ff_exp;
+                            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {    n_embd, n_ff_shexp}, 0);
+                            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd    }, 0);
+                            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, n_ff_shexp}, 0);
+                        } else {
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        }
+                    }
+                } break;
             case LLM_ARCH_DECI:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -1911,6 +2069,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_BERT:
             case LLM_ARCH_NOMIC_BERT:
+            case LLM_ARCH_NOMIC_BERT_MOE:
                 {
                     tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
                     type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0);
@@ -1944,20 +2103,31 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
                         }
 
+                        if (arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                            layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
+                        }
+
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {n_embd, n_embd}, 0);
 
                         layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
                         layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);
 
-                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,        "weight", i), {n_embd, n_ff}, 0);
-                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN,      "weight", i), {n_ff, n_embd}, 0);
-
-                        if (arch == LLM_ARCH_BERT) {
+                        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
                             layer.bo         = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
-                            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, 0);
-                            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
                         } else {
-                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,        "weight", i), {n_embd, n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN,      "weight", i), {n_ff, n_embd}, 0);
+
+                            if (arch == LLM_ARCH_BERT || arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                                layer.bo         = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+                                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, 0);
+                                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
+                            } else {
+                                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+                            }
                         }
 
                         layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
@@ -2210,9 +2380,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
                         // optional bias tensors
-                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, 0);
-                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
-                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -2241,6 +2411,77 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, n_ff_shexp}, 0);
                     }
                 } break;
+            case LLM_ARCH_QWEN3:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                    }
+                } break;
+            case LLM_ARCH_QWEN3MOE:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+                        if (n_expert == 0) {
+                            throw std::runtime_error("n_expert must be > 0 for QWEN3MOE");
+                        }
+                        if (n_expert_used == 0) {
+                            throw std::runtime_error("n_expert_used must be > 0 for QWEN3MOE");
+                        }
+
+                        // MoE branch
+                        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                    }
+                } break;
             case LLM_ARCH_PHI2:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -2329,7 +2570,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
                         layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias",   i), { n_embd }, 0);
 
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, TENSOR_NOT_REQUIRED);
                         if (layer.wqkv == nullptr) {
                             layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
                             layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias",   i), {n_embd}, 0);
@@ -2558,7 +2799,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,   "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -2985,8 +3231,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 {
                     const bool is_lite = (hparams.n_layer == 27);
 
+                    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+                    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+                    const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+                    const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
                     const int64_t n_embd_head_qk_rope = hparams.n_rot;
-                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
 
                     const int64_t q_lora_rank  = hparams.n_lora_q;
                     const int64_t kv_lora_rank = hparams.n_lora_kv;
@@ -3012,14 +3264,22 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                         if (!is_lite) {
                             layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
-                            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}, 0);
+                            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
                         } else {
-                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, 0);
                         }
 
-                        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0);
-                        layer.wkv_b     = create_tensor(tn(LLM_TENSOR_ATTN_KV_B,     "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0);
-                        layer.wo        = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd}, 0);
+                        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
+
+                        // note: only old legacy GGUF files will have the unsplit wkv_b tensor in
+                        if (is_mla) {
+                            layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
+                            layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+                        } else {
+                            layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, 0);
+                        }
+
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -3050,6 +3310,35 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         }
                     }
                 } break;
+            case LLM_ARCH_PLM:
+                {
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const int64_t kv_lora_rank = hparams.n_lora_kv;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    // output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+                    output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq        = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0);
+                        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
+                        layer.wkv_b     = create_tensor(tn(LLM_TENSOR_ATTN_KV_B,     "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0);
+                        layer.wo        = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                    }
+                } break;
             case LLM_ARCH_BITNET:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -3215,16 +3504,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         auto & layer = layers[i];
 
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         if (layer.wqkv == nullptr) {
                             layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
                             layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
                             layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
-                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
-                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
                         }
 
                         layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
@@ -3236,23 +3525,62 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
                     }
                 } break;
-            case LLM_ARCH_NEMOTRON:
+            case LLM_ARCH_GLM4:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     // output
-                    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0);
-                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
 
-                        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
-                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
-                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        }
+
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, 0);
+
+                        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
+            case LLM_ARCH_NEMOTRON:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0);
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
                         layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
@@ -3335,12 +3663,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {time_mix_extra_dim, n_embd, 5}, 0);
 
                         layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0);
-                        layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, TENSOR_NOT_REQUIRED);
                         GGML_ASSERT(!(layer.time_mix_lerp_fused == NULL && layer.time_mix_lerp_w == NULL));
 
                         layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, 0);
@@ -3370,7 +3698,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, TENSOR_NOT_REQUIRED);
                     output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
 
                     const int time_mix_extra_dim = hparams.time_mix_extra_dim;
@@ -3396,7 +3724,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0);
                         layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0);
 
-                        layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, TENSOR_NOT_REQUIRED);
                         layer.time_mix_decay = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY, "weight", i), {n_embd}, 0);
                         layer.time_mix_decay_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W1, "weight", i), {n_embd, time_decay_extra_dim}, 0);
                         layer.time_mix_decay_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W2, "weight", i), {time_decay_extra_dim, attn_hidden_size}, 0);
@@ -3405,9 +3733,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         layer.time_mix_gate = create_tensor(tn(LLM_TENSOR_TIME_MIX_GATE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         // optional bias tensors
-                        layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, TENSOR_NOT_REQUIRED);
 
                         layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
 
@@ -3528,8 +3856,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0);
                         }
 
-                        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, TENSOR_NOT_REQUIRED);
 
                         try {
                             layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0);
@@ -3546,8 +3874,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
 
-                        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
                         layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@@ -3694,6 +4022,46 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     output   = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {hparams.convnext.n_embd, n_embd}, 0);
                     output_b = create_tensor(tn(LLM_TENSOR_OUTPUT, "bias"),   {n_embd}, 0);
                 } break;
+            case LLM_ARCH_BAILINGMOE:
+                {
+                    const int64_t n_ff_exp            = hparams.n_ff_exp;
+                    const int64_t n_expert_shared     = hparams.n_expert_shared;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_head * n_rot}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_head_kv * n_rot}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_head_kv * n_rot}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_rot, n_embd}, 0);
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+                        if (n_expert == 0) {
+                            throw std::runtime_error("n_expert must be > 0");
+                        }
+                        if (n_expert_used == 0) {
+                            throw std::runtime_error("n_expert_used must be > 0");
+                        }
+
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                        layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+                        layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+                        layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -3962,6 +4330,8 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
         LLAMA_LOG_INFO("%s: n_lora_q             = %d\n",     __func__, hparams.n_lora_q);
         LLAMA_LOG_INFO("%s: n_lora_kv            = %d\n",     __func__, hparams.n_lora_kv);
+        LLAMA_LOG_INFO("%s: n_embd_head_k_mla    = %d\n",     __func__, hparams.n_embd_head_k_mla);
+        LLAMA_LOG_INFO("%s: n_embd_head_v_mla    = %d\n",     __func__, hparams.n_embd_head_v_mla);
         LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
         LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
@@ -3975,12 +4345,24 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
     }
 
+    if (arch == LLM_ARCH_QWEN3MOE) {
+        LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
+    }
+
     if (arch == LLM_ARCH_MINICPM || arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
     }
 
+    if (arch == LLM_ARCH_BAILINGMOE) {
+        LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
+        LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
+        LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
+        LLAMA_LOG_INFO("%s: expert_weights_norm  = %d\n",     __func__, hparams.expert_weights_norm);
+    }
+
     vocab.print_info();
 }
 
@@ -4042,6 +4424,10 @@ ggml_backend_buffer_type_t llama_model::select_buft(int il) const {
             });
 }
 
+bool llama_model::has_tensor_overrides() const {
+    return pimpl->has_tensor_overrides;
+}
+
 const ggml_tensor * llama_model::get_tensor(const char * name) const {
     auto it = std::find_if(tensors_by_name.begin(), tensors_by_name.end(),
             [name](const std::pair<std::string, ggml_tensor *> & it) {
@@ -4069,12 +4455,22 @@ struct llm_build_llama : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
+        // temperature tuning
+        ggml_tensor * inp_attn_scale = nullptr;
+        if (arch == LLM_ARCH_LLAMA4) {
+            inp_attn_scale = build_inp_attn_scale();
+        }
+
         auto * inp_attn = build_attn_inp_kv_unified();
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
+            bool use_rope = arch == LLM_ARCH_LLAMA4
+                ? (il + 1) % hparams.n_no_rope_layer_step != 0
+                : true;
+
             // norm
             cur = build_norm(inpL,
                     model.layers[il].attn_norm, NULL,
@@ -4112,25 +4508,38 @@ struct llm_build_llama : public llm_graph_context {
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
                 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
-                Qcur = ggml_rope_ext(
-                        ctx0, Qcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
+                if (use_rope) {
+                    Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, rope_factors,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
 
-                Kcur = ggml_rope_ext(
-                        ctx0, Kcur, inp_pos, rope_factors,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow
-                        );
+                    Kcur = ggml_rope_ext(
+                            ctx0, Kcur, inp_pos, rope_factors,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                } else if (inp_attn_scale) {
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                }
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
+                if (arch == LLM_ARCH_LLAMA4 && use_rope && hparams.use_kq_norm) {
+                    // Llama4TextL2Norm
+                    Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
+                    Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
+                    cb(Qcur, "Qcur_normed", il);
+                    cb(Kcur, "Kcur_normed", il);
+                }
+
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
             }
 
             if (il == n_layer - 1) {
@@ -4148,7 +4557,7 @@ struct llm_build_llama : public llm_graph_context {
             ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
-            // feed-forward network
+            // feed-forward network (non-MoE)
             if (model.layers[il].ffn_gate_inp == nullptr) {
 
                 cur = build_norm(ffn_inp,
@@ -4163,6 +4572,38 @@ struct llm_build_llama : public llm_graph_context {
                         NULL,
                         LLM_FFN_SILU, LLM_FFN_PAR, il);
                 cb(cur, "ffn_out", il);
+
+            } else if (arch == LLM_ARCH_LLAMA4) {
+                // llama4 MoE
+                ggml_tensor * ffn_inp_normed = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, false,
+                        false, 0.0,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
+                        il);
+
+                // Shared experts
+                ggml_tensor * shexp_out = build_ffn(ffn_inp_normed,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(shexp_out, "ffn_moe_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, shexp_out);
+                cb(cur, "ffn_moe_out_merged", il);
+
             } else {
                 // MoE branch
                 cur = build_norm(ffn_inp,
@@ -4310,7 +4751,7 @@ struct llm_build_deci : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -4452,7 +4893,7 @@ struct llm_build_baichuan : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -4567,7 +5008,7 @@ struct llm_build_xverse : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -4692,7 +5133,7 @@ struct llm_build_falcon : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -4822,7 +5263,7 @@ struct llm_build_grok : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -4973,7 +5414,7 @@ struct llm_build_dbrx : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5087,7 +5528,7 @@ struct llm_build_starcoder : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5186,7 +5627,7 @@ struct llm_build_refact : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5313,6 +5754,11 @@ struct llm_build_bert : public llm_graph_context {
                 cur = build_lora_mm(model.layers[il].wqkv, cur);
                 cb(cur, "wqkv", il);
 
+                if (model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
                 Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
                 Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
                 Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
@@ -5340,7 +5786,7 @@ struct llm_build_bert : public llm_graph_context {
 
             cur = build_attn(inp_attn, gf,
                     model.layers[il].wo, model.layers[il].bo,
-                    Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             cb(cur, "kqv_out", il);
 
             if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) {
@@ -5365,13 +5811,29 @@ struct llm_build_bert : public llm_graph_context {
             cb(ffn_inp, "ffn_inp", il);
 
             // feed-forward network
-            if (model.arch == LLM_ARCH_BERT) {
+            if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
+                // MoE branch
+                cur = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        nullptr,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        hparams.n_expert,
+                        hparams.n_expert_used,
+                        LLM_FFN_GELU,
+                        false, false,
+                        0.0f,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+                cb(cur, "ffn_moe_out", il);
+            } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                         NULL,                      NULL,                        NULL,
                         model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                         NULL,
                         LLM_FFN_GELU, LLM_FFN_SEQ, il);
+                cb(cur, "ffn_out", il);
             } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   NULL,                        NULL,
@@ -5379,6 +5841,7 @@ struct llm_build_bert : public llm_graph_context {
                         model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                         NULL,
                         LLM_FFN_GELU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
             } else {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   NULL, NULL,
@@ -5386,8 +5849,8 @@ struct llm_build_bert : public llm_graph_context {
                         model.layers[il].ffn_down, NULL, NULL,
                         NULL,
                         LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
             }
-            cb(cur, "ffn_out", il);
 
             // attentions bypass the intermediate layer
             cur = ggml_add(ctx0, cur, ffn_inp);
@@ -5457,7 +5920,7 @@ struct llm_build_bloom : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5598,7 +6061,7 @@ struct llm_build_mpt : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5744,7 +6207,7 @@ struct llm_build_stablelm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5867,7 +6330,7 @@ struct llm_build_qwen : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5987,7 +6450,7 @@ struct llm_build_qwen2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6108,7 +6571,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6193,16 +6656,25 @@ struct llm_build_qwen2moe : public llm_graph_context {
             {
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
                 cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
 
                 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
                 cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
 
                 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
 
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
@@ -6226,7 +6698,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6257,7 +6729,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
                         false, 0.0,
                         LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
                         il);
-            cb(cur, "ffn_moe_out", il);
+            cb(moe_out, "ffn_moe_out", il);
 
             // FFN shared expert
             {
@@ -6313,16 +6785,14 @@ struct llm_build_qwen2moe : public llm_graph_context {
     }
 };
 
-struct llm_build_phi2 : public llm_graph_context {
-    llm_build_phi2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+struct llm_build_qwen3 : public llm_graph_context {
+    llm_build_qwen3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
 
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         ggml_tensor * cur;
-        ggml_tensor * attn_norm_output;
-        ggml_tensor * ffn_output;
         ggml_tensor * inpL;
 
         inpL = build_inp_embd(model.tok_embd);
@@ -6333,48 +6803,42 @@ struct llm_build_phi2 : public llm_graph_context {
         auto * inp_attn = build_attn_inp_kv_unified();
 
         for (int il = 0; il < n_layer; ++il) {
-            attn_norm_output = build_norm(inpL,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, il);
-            cb(attn_norm_output, "attn_norm", il);
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
 
             // self-attention
             {
-                ggml_tensor * Qcur = nullptr;
-                ggml_tensor * Kcur = nullptr;
-                ggml_tensor * Vcur = nullptr;
-
-                if (model.layers[il].wqkv) {
-                    cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output);
-                    cb(cur, "wqkv", il);
-
-                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                    cb(cur, "bqkv", il);
-
-                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-                } else {
-                    Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq);
-                    Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk);
-                    Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv);
-                }
-
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
                 cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
                 cb(Vcur, "Vcur", il);
 
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
                 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow
                         );
 
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+
                 Kcur = ggml_rope_ext(
                         ctx0, Kcur, inp_pos, nullptr,
                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -6385,36 +6849,36 @@ struct llm_build_phi2 : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                // with phi2, we scale the Q to avoid precision issues
-                // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
-
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
                 // skip computing output for unused tokens
                 ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);
-                inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);
-                attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
-            // FF
-            {
-                ffn_output = build_ffn(attn_norm_output,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
-                cb(ffn_output, "ffn_out", il);
-            }
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
 
-            cur = ggml_add(ctx0, cur, ffn_output);
-            cur = ggml_add(ctx0, cur, inpL);
+            // feed-forward network
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
 
             cur = build_cvec(cur, il);
             cb(cur, "l_out", il);
@@ -6423,18 +6887,17 @@ struct llm_build_phi2 : public llm_graph_context {
             inpL = cur;
         }
 
-        cur = build_norm(inpL,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, -1);
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
 
         cb(cur, "result_norm", -1);
         res->t_embd = cur;
 
+        // lm_head
         cur = build_lora_mm(model.output, cur);
-        cb(cur, "result_output_no_bias", -1);
-
-        cur = ggml_add(ctx0, cur, model.output_b);
 
         cb(cur, "result_output", -1);
         res->t_logits = cur;
@@ -6443,10 +6906,268 @@ struct llm_build_phi2 : public llm_graph_context {
     }
 };
 
-struct llm_build_phi3 : public llm_graph_context {
-    llm_build_phi3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+struct llm_build_qwen3moe : public llm_graph_context {
+    llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out =
+                build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, true,
+                        false, 0.0,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                        il);
+            cb(moe_out, "ffn_moe_out", il);
+            cur = moe_out;
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
+struct llm_build_phi2 : public llm_graph_context {
+    llm_build_phi2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * attn_norm_output;
+        ggml_tensor * ffn_output;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        for (int il = 0; il < n_layer; ++il) {
+            attn_norm_output = build_norm(inpL,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, il);
+            cb(attn_norm_output, "attn_norm", il);
+
+            // self-attention
+            {
+                ggml_tensor * Qcur = nullptr;
+                ggml_tensor * Kcur = nullptr;
+                ggml_tensor * Vcur = nullptr;
+
+                if (model.layers[il].wqkv) {
+                    cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output);
+                    cb(cur, "wqkv", il);
+
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+
+                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                } else {
+                    Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq);
+                    Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk);
+                    Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv);
+                }
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                // with phi2, we scale the Q to avoid precision issues
+                // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
+                Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);
+                inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);
+                attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);
+            }
+
+            // FF
+            {
+                ffn_output = build_ffn(attn_norm_output,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, il);
+                cb(ffn_output, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_output);
+            cur = ggml_add(ctx0, cur, inpL);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = build_norm(inpL,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        cur = build_lora_mm(model.output, cur);
+        cb(cur, "result_output_no_bias", -1);
+
+        cur = ggml_add(ctx0, cur, model.output_b);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
+struct llm_build_phi3 : public llm_graph_context {
+    llm_build_phi3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
 
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
@@ -6520,7 +7241,7 @@ struct llm_build_phi3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -6655,7 +7376,7 @@ struct llm_build_plamo : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             ggml_tensor * sa_out = cur;
 
@@ -6762,7 +7483,7 @@ struct llm_build_gpt2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6878,7 +7599,7 @@ struct llm_build_codeshell : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7007,7 +7728,7 @@ struct llm_build_orion : public llm_graph_context {
 
             cur = build_attn(inp_attn, gf,
                     model.layers[il].wo, NULL,
-                    Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
         }
 
         if (il == n_layer - 1) {
@@ -7134,7 +7855,7 @@ struct llm_build_internlm2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7331,7 +8052,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -7461,7 +8182,7 @@ struct llm_build_gemma : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -7583,7 +8304,7 @@ struct llm_build_gemma2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             cur = build_norm(cur,
@@ -7724,7 +8445,7 @@ struct llm_build_gemma3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, hparams.f_attention_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
             }
 
             cur = build_norm(cur,
@@ -7864,7 +8585,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8199,7 +8920,7 @@ struct llm_build_command_r : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8334,7 +9055,7 @@ struct llm_build_cohere2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8465,7 +9186,7 @@ struct llm_build_olmo : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8585,7 +9306,7 @@ struct llm_build_olmo2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             cur = build_norm(cur,
@@ -8718,7 +9439,7 @@ struct llm_build_olmoe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8851,7 +9572,7 @@ struct llm_build_openelm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8965,7 +9686,7 @@ struct llm_build_gptneox : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9115,7 +9836,7 @@ struct llm_build_arctic : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9270,7 +9991,7 @@ struct llm_build_deepseek : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -9360,16 +10081,23 @@ struct llm_build_deepseek2 : public llm_graph_context {
     llm_build_deepseek2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         bool is_lite = (hparams.n_layer == 27);
 
-        // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
-        // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
-        const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
-        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k));
-        const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
+        const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+        // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+        const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+        const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
+        const int64_t n_embd_head_qk_rope = hparams.n_rot;
+        const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
 
-        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
         const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
+        // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+        // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+        const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(n_embd_head_k));
+        const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
+
         ggml_tensor * cur;
         ggml_tensor * inpL;
 
@@ -9394,16 +10122,14 @@ struct llm_build_deepseek2 : public llm_graph_context {
             {
                 ggml_tensor * q = NULL;
                 if (!is_lite) {
-                    // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
                     q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
                     cb(q, "q", il);
 
                     q = build_norm(q,
-                            model.layers[il].attn_q_a_norm, NULL,
+                            model.layers[il].attn_q_a_norm, nullptr,
                             LLM_NORM_RMS, il);
                     cb(q, "q", il);
 
-                    // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
                     q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
                     cb(q, "q", il);
                 } else {
@@ -9411,96 +10137,125 @@ struct llm_build_deepseek2 : public llm_graph_context {
                     cb(q, "q", il);
                 }
 
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                ggml_tensor * q_nope = ggml_view_3d(ctx0, q,
+                        n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(q->type, n_embd_head_k),
+                        ggml_row_size(q->type, n_embd_head_k) * n_head,
                         0);
                 cb(q_nope, "q_nope", il);
 
-                // and {n_head * n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                // and {n_embd_head_qk_rope, n_head, n_tokens}
+                ggml_tensor * q_pe = ggml_view_3d(ctx0, q,
+                        n_embd_head_qk_rope, n_head, n_tokens,
+                        ggml_row_size(q->type, n_embd_head_k),
+                        ggml_row_size(q->type, n_embd_head_k) * n_head,
                         ggml_row_size(q->type, n_embd_head_qk_nope));
                 cb(q_pe, "q_pe", il);
 
-                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
-                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+                ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+                cb(kv_cmpr_pe, "kv_cmpr_pe", il);
 
                 // split into {kv_lora_rank, n_tokens}
-                ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
-                        kv_pe_compresseed->nb[1],
+                ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe,
+                        kv_lora_rank, n_tokens,
+                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
                         0);
-                cb(kv_compressed, "kv_compressed", il);
+                cb(kv_cmpr, "kv_cmpr", il);
+
+                // and {n_embd_head_qk_rope, 1, n_tokens}
+                ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe,
+                        n_embd_head_qk_rope, 1, n_tokens,
+                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+                cb(k_pe, "k_pe", il);
 
-                // and {n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        kv_pe_compresseed->nb[1],
-                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+                q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(q_pe, "q_pe", il);
+
+                k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                );
                 cb(k_pe, "k_pe", il);
 
-                // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont
-                kv_compressed = ggml_cont(ctx0, kv_compressed);
-                kv_compressed = build_norm(kv_compressed,
-                        model.layers[il].attn_kv_a_norm, NULL,
+                kv_cmpr = build_norm(kv_cmpr,
+                        model.layers[il].attn_kv_a_norm, nullptr,
                         LLM_NORM_RMS, il);
-                cb(kv_compressed, "kv_compressed", il);
+                cb(kv_cmpr, "kv_cmpr", il);
 
-                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
-                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
-                cb(kv, "kv", il);
+                if (is_mla) {
+                    // {n_embd_head_qk_nope, n_tokens, n_head}
+                    q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                    cb(q_nope, "q_nope_perm", il);
 
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        0);
-                cb(k_nope, "k_nope", il);
+                    // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+                    ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                    cb(q_nope_absorbed, "q_nope_absorbed", il);
 
-                // and {n_head * n_embd_head_v, n_tokens}
-                ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
-                cb(v_states, "v_states", il);
+                    // {kv_lora_rank, n_head, n_tokens}
+                    q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                    cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
 
-                v_states = ggml_cont(ctx0, v_states);
-                cb(v_states, "v_states", il);
+                    // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+                    // note: rope must go first for in-place context shifting in build_rope_shift()
+                    ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                    cb(Qcur, "Qcur", il);
 
-                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
-                        ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
-                        0);
-                cb(v_states, "v_states", il);
+                    kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                    cb(kv_cmpr, "kv_cmpr_reshape", il);
 
-                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                q_pe = ggml_rope_ext(
-                        ctx0, q_pe, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor_scaled, beta_fast, beta_slow
-                        );
-                cb(q_pe, "q_pe", il);
+                    // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                    ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                    cb(Kcur, "Kcur", il);
 
-                // shared RoPE key
-                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                k_pe = ggml_rope_ext(
-                        ctx0, k_pe, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor_scaled, beta_fast, beta_slow
-                        );
-                cb(k_pe, "k_pe", il);
+                    // {kv_lora_rank, 1, n_tokens}
+                    ggml_tensor * Vcur = kv_cmpr;
+                    cb(Vcur, "Vcur", il);
 
-                ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
-                cb(q_states, "q_states", il);
+                    // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
+                    cur = build_attn(inp_attn, gf,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, model.layers[il].wv_b, kq_scale, il);
+                } else {
+                    ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
+                    cb(kv, "kv", il);
+
+                    // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                    ggml_tensor * k_nope = ggml_view_3d(ctx0, kv,
+                            n_embd_head_qk_nope, n_head, n_tokens,
+                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                            0);
+                    cb(k_nope, "k_nope_view", il);
 
-                ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
-                cb(k_states, "k_states", il);
+                    // and {n_embd_head_v, n_head, n_tokens}
+                    ggml_tensor * Vcur = ggml_view_3d(ctx0, kv,
+                            n_embd_head_v, n_head, n_tokens,
+                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                            ggml_row_size(kv->type, n_embd_head_qk_nope));
+                    cb(Vcur, "Vcur_view", il);
 
-                cur = build_attn(inp_attn, gf,
-                        model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, kq_scale, il);
+                    Vcur = ggml_cont(ctx0, Vcur);
+                    cb(Vcur, "Vcur_cont", il);
+
+                    // note: rope must go first for in-place context shifting in build_rope_shift()
+                    ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
+                    cb(Qcur, "Qcur", il);
+
+                    ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                    cb(Kcur, "Kcur", il);
+
+                    // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
+                    cur = build_attn(inp_attn, gf,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                }
             }
 
             if (il == n_layer - 1) {
@@ -9666,7 +10421,7 @@ struct llm_build_bitnet : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         NULL, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 
                 cur = build_norm(cur,
                         model.layers[il].attn_sub_norm, NULL,
@@ -9789,7 +10544,7 @@ struct llm_build_t5_enc : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo_enc, nullptr,
-                        Qcur, Kcur, Vcur, kq_b, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -9895,7 +10650,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
                 cur = build_attn(inp_attn_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, kq_b, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -9927,7 +10682,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
                 cur = build_attn(inp_attn_cross, gf,
                         model.layers[il].wo_cross, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
 
                 //ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
@@ -10060,7 +10815,7 @@ struct llm_build_jais : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/float(n_embd_head), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il);
             }
 
             if (il == n_layer - 1) {
@@ -10192,7 +10947,7 @@ struct llm_build_chatglm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -10245,6 +11000,157 @@ struct llm_build_chatglm : public llm_graph_context {
     }
 };
 
+struct llm_build_glm4 : public llm_graph_context {
+    llm_build_glm4(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // Pre-attention norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                ggml_tensor * Qcur = nullptr;
+                ggml_tensor * Kcur = nullptr;
+                ggml_tensor * Vcur = nullptr;
+
+                if (model.layers[il].wqkv == nullptr) {
+                    Qcur = build_lora_mm(model.layers[il].wq, cur);
+                    if (model.layers[il].bq) {
+                        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    }
+                    Kcur = build_lora_mm(model.layers[il].wk, cur);
+                    if (model.layers[il].bk) {
+                        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    }
+                    Vcur = build_lora_mm(model.layers[il].wv, cur);
+                    if (model.layers[il].bv) {
+                        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    }
+                } else {
+                    cur = build_lora_mm(model.layers[il].wqkv, cur);
+                    cb(cur, "wqkv", il);
+                    if (model.layers[il].bqkv) {
+                        cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                        cb(cur, "bqkv", il);
+                    }
+                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // Post-attention norm (new!)
+            cur = build_norm(cur,
+                    model.layers[il].attn_post_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "post_attn_norm", il);
+
+            // Add the input (residual connection after post-attention norm)
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                // Pre-MLP norm
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm,
+                        NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                // MLP
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        NULL,                      NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+                cb(cur, "ffn_out", il);
+
+                // Post-MLP norm
+                cur = build_norm(cur,
+                        model.layers[il].ffn_post_norm,
+                        NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "post_mlp_norm", il);
+            }
+
+            // Add residual connection after post-MLP norm
+            inpL = ggml_add(ctx0, cur, ffn_inp);
+            cb(inpL, "l_out", il);
+        }
+
+        // Final norm
+        cur = build_norm(inpL,
+                model.output_norm,
+                NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // Output projection
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 struct llm_build_nemotron : public llm_graph_context {
     llm_build_nemotron(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -10318,7 +11224,7 @@ struct llm_build_nemotron : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -10449,7 +11355,7 @@ struct llm_build_exaone : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -11351,7 +12257,7 @@ struct llm_build_chameleon : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 
                 if (hparams.swin_norm) {
                     cur = build_norm(cur,
@@ -11588,32 +12494,348 @@ struct llm_build_wavtokenizer_dec : public llm_graph_context {
     }
 };
 
-llama_memory_i * llama_model::create_memory() const {
-    llama_memory_i * res;
+struct llm_build_plm : public llm_graph_context {
+    llm_build_plm(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k));
 
-    switch (arch) {
-        case LLM_ARCH_MAMBA:
-        case LLM_ARCH_RWKV6:
-        case LLM_ARCH_RWKV6QWEN2:
-        case LLM_ARCH_RWKV7:
-        case LLM_ARCH_ARWKV7:
-            {
-                res = new llama_kv_cache_unified(hparams, {
-                    /*.get_rope_factors =*/ nullptr
-                });
-            } break;
-        default:
-            {
-                res = new llama_kv_cache_unified(hparams, {
-                    /*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) {
-                        // choose long/short freq factors based on the context size
-                        if (layers[il].rope_freqs != nullptr) {
-                            return layers[il].rope_freqs;
-                        }
+        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
+        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+        const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
-                        if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
-                            return layers[il].rope_long;
-                        }
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        // {n_embd, n_tokens}
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                ggml_tensor * q = NULL;
+                q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(q, "q", il);
+
+                // split into {n_head * n_embd_head_qk_nope, n_tokens}
+                ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(q->type, hparams.n_embd_head_k),
+                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                        0);
+                cb(q_nope, "q_nope", il);
+
+                // and {n_head * n_embd_head_qk_rope, n_tokens}
+                ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                        ggml_row_size(q->type, hparams.n_embd_head_k),
+                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                        ggml_row_size(q->type, n_embd_head_qk_nope));
+                cb(q_pe, "q_pe", il);
+
+                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
+                ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+
+                // split into {kv_lora_rank, n_tokens}
+                ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
+                        kv_pe_compresseed->nb[1],
+                        0);
+                cb(kv_compressed, "kv_compressed", il);
+
+                // and {n_embd_head_qk_rope, n_tokens}
+                ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
+                        kv_pe_compresseed->nb[1],
+                        kv_pe_compresseed->nb[1],
+                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+                cb(k_pe, "k_pe", il);
+
+                kv_compressed = build_norm(kv_compressed,
+                        model.layers[il].attn_kv_a_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(kv_compressed, "kv_compressed", il);
+
+                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
+                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
+                cb(kv, "kv", il);
+
+                // split into {n_head * n_embd_head_qk_nope, n_tokens}
+                ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
+                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                        0);
+                cb(k_nope, "k_nope", il);
+
+                // and {n_head * n_embd_head_v, n_tokens}
+                ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
+                cb(v_states, "v_states", il);
+
+                v_states = ggml_cont(ctx0, v_states);
+                cb(v_states, "v_states", il);
+
+                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
+                        ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
+                        0);
+                cb(v_states, "v_states", il);
+
+                q_pe = ggml_rope_ext(
+                        ctx0, q_pe, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+                cb(q_pe, "q_pe", il);
+
+                // shared RoPE key
+                k_pe = ggml_rope_ext(
+                        ctx0, k_pe, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+                cb(k_pe, "k_pe", il);
+
+                ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+                cb(q_states, "q_states", il);
+
+                ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+                cb(k_states, "k_states", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, NULL,
+                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
+struct llm_build_bailingmoe : public llm_graph_context {
+    llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out =
+                build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, hparams.expert_weights_norm,
+                        false, hparams.expert_weights_scale,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                        il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
+llama_memory_i * llama_model::create_memory() const {
+    llama_memory_i * res;
+
+    switch (arch) {
+        case LLM_ARCH_MAMBA:
+        case LLM_ARCH_RWKV6:
+        case LLM_ARCH_RWKV6QWEN2:
+        case LLM_ARCH_RWKV7:
+        case LLM_ARCH_ARWKV7:
+            {
+                res = new llama_kv_cache_unified(hparams, {
+                    /*.get_rope_factors =*/ nullptr
+                });
+            } break;
+        default:
+            {
+                res = new llama_kv_cache_unified(hparams, {
+                    /*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) {
+                        // choose long/short freq factors based on the context size
+                        if (layers[il].rope_freqs != nullptr) {
+                            return layers[il].rope_freqs;
+                        }
+
+                        if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
+                            return layers[il].rope_long;
+                        }
 
                         return layers[il].rope_short;
                     }
@@ -11632,6 +12854,7 @@ llm_graph_result_ptr llama_model::build_graph(
 
     switch (arch) {
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_LLAMA4:
         case LLM_ARCH_MINICPM:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
@@ -11665,6 +12888,7 @@ llm_graph_result_ptr llama_model::build_graph(
         case LLM_ARCH_BERT:
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
             {
                 llm = std::make_unique<llm_build_bert>(*this, params, gf);
             } break;
@@ -11696,6 +12920,14 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_qwen2moe>(*this, params, gf);
             } break;
+        case LLM_ARCH_QWEN3:
+            {
+                llm = std::make_unique<llm_build_qwen3>(*this, params, gf);
+            } break;
+        case LLM_ARCH_QWEN3MOE:
+            {
+                llm = std::make_unique<llm_build_qwen3moe>(*this, params, gf);
+            } break;
         case LLM_ARCH_PHI2:
             {
                 llm = std::make_unique<llm_build_phi2>(*this, params, gf);
@@ -11801,6 +13033,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_chatglm>(*this, params, gf);
             } break;
+        case LLM_ARCH_GLM4:
+            {
+                llm = std::make_unique<llm_build_glm4>(*this, params, gf);
+            } break;
         case LLM_ARCH_BITNET:
             {
                 llm = std::make_unique<llm_build_bitnet>(*this, params, gf);
@@ -11819,10 +13055,11 @@ llm_graph_result_ptr llama_model::build_graph(
                         GGML_ABORT("invalid graph type");
                 };
             } break;
-            //case LLM_ARCH_T5ENCODER:
-            //    {
-            //        llm.build_t5_enc(gf);
-            //    } break;
+        case LLM_ARCH_T5ENCODER:
+            {
+                llm = std::make_unique<llm_build_t5_enc>(*this, params, gf);
+            }
+            break;
         case LLM_ARCH_JAIS:
             {
                 llm = std::make_unique<llm_build_jais>(*this, params, gf);
@@ -11859,6 +13096,14 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params, gf);
             } break;
+        case LLM_ARCH_PLM:
+            {
+                llm = std::make_unique<llm_build_plm>(*this, params, gf);
+            } break;
+        case LLM_ARCH_BAILINGMOE:
+            {
+                llm = std::make_unique<llm_build_bailingmoe>(*this, params, gf);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -11876,6 +13121,7 @@ llm_graph_result_ptr llama_model::build_graph(
 llama_model_params llama_model_default_params() {
     llama_model_params result = {
         /*.devices                     =*/ nullptr,
+        /*.tensor_buft_overrides       =*/ nullptr,
         /*.n_gpu_layers                =*/ 0,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
@@ -11971,6 +13217,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_LLAMA4:
         case LLM_ARCH_DECI:
         case LLM_ARCH_BAICHUAN:
         case LLM_ARCH_STARCODER:
@@ -11985,10 +13232,13 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARCTIC:
         case LLM_ARCH_DEEPSEEK:
         case LLM_ARCH_DEEPSEEK2:
+        case LLM_ARCH_PLM:
         case LLM_ARCH_CHATGLM:
+        case LLM_ARCH_GLM4:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_CHAMELEON:
+        case LLM_ARCH_BAILINGMOE:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -11997,11 +13247,14 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DBRX:
         case LLM_ARCH_BERT:
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_STABLELM:
         case LLM_ARCH_BITNET:
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_QWEN2MOE:
+        case LLM_ARCH_QWEN3:
+        case LLM_ARCH_QWEN3MOE:
         case LLM_ARCH_OLMO2:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2: